On the Performance of Virtualized Infrastructures for Processing Realtime Streaming Data

Abstract

Clouds have become ubiquitous and several data processing tasks have migrated to these settings. The dominant approach in cloud settings is to provision virtual machines (VMs) rather than provision direct access to the physical machine. One artifact of such provisioning is that multiple VMs may be collocated on the same physical machine and possibly interfere with each other. In this paper, we focus on the impact of virtualized infrastructures on real time stream processing, we use the classification of electrocardiograms (ECG) as a motivating example. Stream processing in such a setting strains resources differently than the traditional web services or analytics on large datasets traditionally performed in the cloud. In streaming environments all processing per packet needs to be completed in a timely manner, and the number and rate at which these packets are generated is high. Our focus is to study the implications of various combinations of virtualization strategies on the performance of real time stream processing. We have done extensive performance benchmarks (using Xen and KVM) the results of which form the basis for our recommendations for the trade-offs involved in these settings.